Linking Caplet and Swaption Volatilities in a BGM/J Framework: Approximate Solutions

Peter J ckel a Riccardo Rebonato Quantitative Research Centre The Royal Bank of Scotland 135 Bishopsgate, London EC2M 3UR August 11th , 2000

Abstract We present and approximation for the volatility of European swaptions in a forward rate based Brace-Gatarek-Musiela/Jamshidian framework [BGM97, Jam97] which enables us to calculate prices for swaptions without the need for Monte Carlo simulations. Also, we explain the mechanism behind the remarkable accuracy of these approximate prices. For cases where the yield curve varies noticeably as a function of maturity, a second, and even more accurate formula is derived.

1 Introduction and motivation

In a forward-rate based BGM/J [BGM97, Jam97] approach, once the time-dependent instantaneous volatilities and correlations of the forward rates have been specied, their stochastic evolution is completely determined. Since swap rates are linear combinations (with stochastic weights) of forward rates, it follows that their dynamics are also fully determined once the volatilities and correlations of the forward rates have been specied. Some (very rare) complex derivatives depend exclusively on the volatility of either set of state variables (forward or swap rates). In general, one set of rates dominate the value of a given product, but the other set still contribute to a signicant extent. Trigger swaps are a classic example of a product where the relative location of the strike and the barrier level can radically shift the relative importance of forward and swap rates. In practical applications it is therefore extremely important to ascertain the implications for the dynamics of the swap rates, given a particular choice of dynamics for the forward rates and vice versa. Unfortunately, as shown later on, the correct evaluation of the swaption

It is a pleasure to thank Dr. Joshi and Dr. Hunter for useful discussions.

prices implied by a choice of forward rate volatilities and correlations is a conceptually straightforward, but computationally rather intensive exercise. This note therefore presents two simple but very effective approximations which allow the estimation of a full swaption volatility matrix in a fraction of a second.

2 Statement of the problem

We begin by placing ourselves in a forward-rate-based BGM/J framework. As mentioned above, the noarbitrage evolution of the forward rates is specied by the choice of a particular functional form for the forward-rate instantaneous volatilities and for the forward-rate/forward-rate correlation function. The task at this point is to obtain the corresponding swap-rate instantaneous volatilities. Let N M (t) denote the relative instantaneous volatility at time t of a swap rate SRN M expiring N years from today and maturing M years thereafter. This swap rate can be viewed as depending on the forward rates of that part of the yield curve in an approximately linear way, namely
n

SRN M (t) =
i=1

wi fi (t) ,

(1)

with the weights wi given by wi = Pi+1 i

n

(2)

Pj+1 j
k=1

In equation (2), Pi+1 denotes the zero coupon bond maturing at the payment time of the i-th forward rate fi , i is the associated accrual factor such that
i 1

Pi+1 =
k=1

(1 + fk k )

P1 ,

(3)

n is the number of forward rates in the swap as illustrated schematically in gure 1, and we have identi-

f1
today

fi ti i tn

fn t n+1

t1
Zero coupon bond P i+1

Figure 1: The forward rates determining the individual payments of the swap.

ed t1 := N and tn+1 := N + M . A straightforward application of It s lemma gives1 o

n

wj wk fj fk jk j k
N M 2

j,k=1 n 2

(4)

wi fi
i=1

where dependence on time has been omitted for clarity and, as usual, j (t) is the time-t instantaneous volatility of log-normal forward rate fj , and jk (t) is the instantaneous correlation between forward rate fj and fk . Expression (4) shows that the instantaneous volatility at time t > 0 of a swap rate is a stochastic quantity, depending as it does on the coefcients {w}, and on the future realization of the forward rates underlying the swap rate {f }. Insofar as the weights {w} are concerned, which are functions of discount factors, one might be tempted to claim that their volatility should be very low compared to that of the swap or forward rates, and, as such, negligible. The same argument, however, certainly cannot be made about the forward rates themselves that enter equation (4). One therefore reaches the conclusion that, starting from a purely deterministic function of time for the instantaneous volatilities of the forward rates, one arrives at a rather complex, and stochastic, expression for the instantaneous volatility of the corresponding swap rate. Therefore, in order to obtain the price of a European swaption corresponding to a given choice of forward-rate instantaneous volatilities, one is faced with a computationally rather cumbersome task: to begin with, in order to obtain the total Black volatility of a given European swaption to expiry, in fact, one rst has to integrate its swap-rate instantaneous volatility
N M 2 Black t1

t1

=
u=0

N M (u) du

(5)

with t1 being the time horizon of expiry of the option in N years from today as dened before. As equation (4) shows, however, at any time u there is one (different) swap rate instantaneous volatility for any future realization of the forward rates from today to time u. But since every path gives rise to a particular swap-rate instantaneous volatility via the dependence on the path of the quantities {w} and {f }, there seems to be no such thing as a single unique total Black volatility for the swap rate. Rather, if one starts from a description of the dynamics of forward rates in terms of a deterministic volatility, there appears to be one Black volatility for a given swap rate associated with each and every realization of the forward rates along the path of the integral. Notice that the implications of equation (4) are farther-reaching than the usual (and correct) claim that log-normal forward rates are incompatible with log-normal swap rates2 . By equation (4) one can conclude that, starting from a purely deterministic volatility for the (logarithm of) the forward rates, the instantaneous volatility of the corresponding swap rate is a stochastic quantity, and that the quantity
1

t1 u=0

N M (u) du is a path-dependent integral that

2

N M cannot be equated to the (path-independent) real number Black

t1 . Calculating the value of several

This statement is not strictly correct. Equation (4) is based on the approximation that the weights wi dont depend on the forward rates which they, in reality, do according to equation (3). See section 3 for a more precise discussion. 2 See [Reb99] for a discussion of the price implications of the joint log-normal assumptions.

European swaptions, or, perhaps, of the whole swaption matrix, therefore becomes a very burdensome task, the more so if the coefcients of the forward-rate instantaneous volatilities are not given a priori but are to be optimised via a numerical search procedure so as to produce, say, the best possible t to the swaption market. Some very simple but useful approximations are however possible. In order to gain some insight into the structure of equation (4), one can begin by regarding it as a weighted average of the products jk (t)j (t)k (t) with doubly-indexed coefcients jk (t) given by jk (t) = wj (t)fj (t) wk (t)fk (t)
n 2

(6)

wi (t)fi (t)
i=1

to convert equation (4) to

n

N M

(t)

=
j,k=1

jk (t)jk (t)j (t)k (t) .

(7)

For a given point in time, and for a given realization of the forward rates, these coefcients are, in general, far from constant or deterministic. Their stochastic evolution is fully determined by the evolution of the forward rates. One can begin, though, to distinguish two important cases: the rst (case 1) refers to (proportionally) parallel moves in the yield curve; the second (case 2) occurs when the yield curve experiences more complex changes. Intuitively, one can therefore regard the results presented in the following as pertaining to movements of the yield curve under shocks of the rst principal component3 for case 1, or to higher principal components for case 2. For the purpose of the discussion to follow, it is also important to keep in mind the typical relative magnitude of the rst principal component shocks relative to the higher modes of deformation. With this distinction in mind, one can rst of all notice that in the rst (parallel) case each individual weight is only mildly dependent on the stochastic realization of the forward rate at time t. Intuitively this can be understood by observing that a given forward rate occurs both in the numerator and in the denominator of equation (6). So the effects on the coefcients of a (reasonably small) identical proportional change in the forward rates to a large extent cancel out. This is shown in gures 2, 3, and 4 for the particular case study illustrated in table 1. The rst of the three gures shows the changes to which the yield curve was subjected (rigid up and down shifts by 25 basis points); gure 3 displays the percentage changes in the coefcients {} for the longest co-terminal swap in moving from the initial yield curve to the yield curve shocked upwards by 25 basis points; gure 4 then shows the average of the percentage changes in the coefcients {} corresponding to the equi-probable up and down 25 basis point shifts. For more complex changes in the shape of the yield curve, the individual coefcients remain less and less constant with increasing order of the principal component. In the less benign case of tilts and bends in the forward curve, the difference between the coefcients calculated with the initial values of
Since, as shown below, the coefcients are approximately constant for identical proportional changes in the forward rates, the principal components should be thought of as referring to log changes.
3

ti 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75

Pi 1.000000 0.986976 0.972996 0.958196 0.942707 0.926657 0.910169 0.893360 0.876340 0.859220 0.842080 0.825040 0.808160 0.791520 0.775210 0.759280 0.743790 0.728800 0.714360 0.700510 0.687290 0.674740 0.662900 0.651800 0.641470

fi 5.278% 5.747% 6.178% 6.572% 6.928% 7.246% 7.526% 7.768% 7.972% 8.138% 8.265% 8.355% 8.406% 8.419% 8.393% 8.329% 8.227% 8.087% 7.908% 7.692% 7.437% 7.144% 6.814% 6.445%

wi

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 5.25 5.50 5.75 6.00

0.072409 0.070965 0.069529 0.068106 0.066704 0.065329 0.063987 0.062681 0.061418 0.060201 0.059034 0.057920 0.056863 0.055865 0.054929 0.054058

Table 1: The initial yield curve and the column of weights {w} for the 2 4-quarterly swaption. the forward rates and after the yield curve move will in general be signicant. However, in these cases one observes that the average of each individual weight corresponding to a positive and negative move of the same magnitude (clockwise and counter-clockwise tilt, increased and decreased curvature) is still remarkably constant. This feature, needless to say, is even more marked for the parallel movement, as shown in gure 4. On the basis of these observations we are therefore in a position to reach two simple but useful conclusions: 1. To the extent that the movements in the forward curve are dominated by a rst (parallel) principal component, the coefcients {} are only very mildly dependent on the path realizations. 2. Even if higher principal components are allowed to shock the forward curve, the expectation of the future swap rate instantaneous volatility is very close to the value obtainable by using todays values for the coefcients {} and the forward rates {f }. Note that the second statement has wider applicability (it does not require that the forward curve should only move in parallel), but yields weaker results, only referring as it does to the average of the instanta5

Figure 2: Yield curves (rigid shift) neous volatility. Note also that the average of the weights over symmetric shocks becomes less and less equal to the original weights as the complexity of the deformation increases; on the other hand we know that relatively few principal components can describe the yield curve dynamics to a high degree of accuracy. Therefore, the negative impact of a progressively poorer approximation becomes correspondingly smaller and smaller. Given the linearity of the integration operator which allows the user to move from instantaneous volatilities to Black volatilities via equation (5), the second conclusion can be transferred to these latter quantities, and it can therefore be re-stated as: Even if higher principal components are allowed to shock the forward curve, the expectation of the average Black volatility is very close to the value obtainable by integrating the swap rate instantaneous volatilities calculated using todays values for the coefcients {} and the forward rates {f }. It is well known, on the other hand, that the price of an at-the-money plain-vanilla option, such as a European swaption, is to a very good approximation a linear function of its implied Black volatility4 . Therefore it follows that, as long as one includes in the average all possible changes in shape of the forward curve in a symmetric fashion, the resulting average of the prices for the European swaption under study obtained using the different weights, jk (t), will be very close to the single price obtained using the current values for {} and {f }.
4

See e.g. [LS95].

hyw x vtqrpghfd us i e ca b S R

Q I PH G E F3D C A B3@ 9 7 836 5 2 431 0 ( )$' & # %$" !        `

Figure 3: Percentage differences in the elements of the coefcients caused by a 25 basis points upward shift of the yield curve This conclusion, by itself, would not be sufcient to authorize the trader to quote as the price for the European swaption the (approximate) average over the price distribution. More precisely, the situation faced by a trader who uses a forward-rate-based implementation for pricing and hedging is as follows: starting from a dynamics for forward rates described by a deterministic volatility, he arrives at a distribution of swaption prices. Therefore, by engaging in a self-nancing trading strategy in forward rates to hedge a swaption assunimg i) that both sets of quantities are log-normally distributed and ii) that. their volatilities are simultaneously deterministic, he will not, in general, manage to produce an exact replication of the swaption payoff by its expiry; therefore the combined portfolio (swaption plus dynamically re-hedged holdings of forward rates) will have a nite variance (and higher moments) at expiry. Strictly speaking, the risk-averse trader will therefore not make a price simply by averaging over the nal portfolio outcomes. The dispersion of the swaption prices around their average is however very small. If one therefore assumes that swaptions and forward rates can have simultaneously deterministic volatilities, and makes use of the results in [Reb98] about the likely impact of the joint log-normal assumption, it is possible to engage in a trading strategy that will produce, by expiry, imperfect but very good replication. In other terms, the trader who were to associate to the swaption a price signicantly different from the

   

%! $ #"

320)!& 1 ( '

9207!4 8 6 5

D BA C@

I GF HE

V U 2TSQRP
7

aYW ` X

Figure 4: Average percentage differences in the elements of the coefcients caused by equi-probable 25 basis points upward and downward shifts of the yield curve average would have to have a utility function exceedingly sensitive indeed to small variations in his nal wealth. Therefore statements 1 and 2 together lead one to surmise that the expression
n

should yield a useful approximation to the instantaneous volatility of the swap rate, and, ultimately, to the European swaption price. It is essential to note that the above equation differs subtly but fundamentally from equation (7) in that the coefcients {} are no longer stochastic quantities, but are evaluated using todays known values for the forward rates and discount factors. Whats more, by virtue of the previous results on the average of the coefcients, a robust approximation for the equivalent implied Black volatility of a European swaption can be derived since the risk-neutral price of an option is given by the expectation, i.e. the average over the risk-neutral measure. The expression for the average Black

    

# "  !

301)(&$ 2 ' %

9 8 6 5 7(&4

F E B A G(DC&@

R Q I DPH

W V G(DUTS

N M

b`(X a Y
(t)
2 j,k=1

jk (0)jk (t)j (t)k (t)

(8)

volatility then becomes

n N M Black 1

=
j,k=1

jk (0)

j (ut1 )k (ut1 )jk (ut1 ) u=0

du

(9)

Equation (9) should be very useful in the context of calibration of FRA-based BGM/J models to market given European swaption volatilities. It enables us to calculate prices for the whole swaption matrix without having to carry out a single Monte Carlo simulation and thus to solve the highly cumbersome problem of calibration with great ease. As shown in the result section, the quality of this approximation is very good. In those situations (noticeably non-at yield curves) where it begins to prove unsatisfactory, it can be easily improved upon by a natural extension, which is presented in the next section.

Rening the approximation

The application of It s lemma to equation (1) actually gives equation (4) only if one assumes that the o weights {w} are independent of the forward rates {f }. More correctly, and neglecting the deterministic part irrelevant for this discussion5 , It s lemma gives o dSR = SR
n

i=1 n

SR dfi fi SR SR fi i dWi fi SR

=
i=1

(10)

wherein the Wiener processes Wi are correlated, i.e. dWi dWj = ij dt . Given the denition
n

(11)

Ai =
j=i

Pj+1 fj j

(12)

of co-terminal oating-leg values and

n

Bi =
j=i

Pj+1 j

(13)

for the co-terminal xed-leg annuities, we obtain after some algebraic manipulations SR = fi
5

Pi+1 i i Ai i A1 Bi + 2 B1 1 + fi i B1 1 + fi i B1

(14)

The neglected terms are truly irrelevant in the following since they drop out as soon as we calculate instantaneous swaprate/swap-rate covariances.

This enables us to calculate the following improved formula for the coefcients {} : ij

P f (A1 Bi Ai B1 )fi i Pj+1 fj j (A1 Bj Aj B1 )fj j i+1 i i = + + A1 A1 B1 (1 + fi i ) A1 A1 B1 (1 + fj j )

as in equation (6) shape correction as in equation (6) shape correction

(15)

We call the second term inside the square brackets of equation (15) the shape correction. Rewriting this corrective term as (A1 Bi Ai B1 )fi i fi i = A1 B1 (1 + fi i ) A1 B1 (1 + fi i )
i1 n

Pl+1 Pm+1 l m (fl fm )

l=1 m=i

(16)

highlights that it is a weighted average over inhomogeneities of the yield curve. In fact, for a at yield curve, all of the terms (fl fm ) are identically zero and the righ-hand-side of equation (15) is identical to that of equation (6).

Specic functional forms

We are nally in a position to conduct some empirical tests. In order to do so, however, one needs to specify a correlation function jk . In general this function will depend both on calendar time, and on the expiry time of the two forward rates. If one makes the assumptions i) that the correlation function is time homogeneous, and ii) that it only depends on the relative distance in years between the two forward rates in question (i.e. on |tj tk |), further simplications are possible. The expression for the average Black volatility now becomes:
n N M 2 Black t1

t1

=
j,k=1

jk (0) |tj tk |

j (u)k (u) u=0

du

(17)

Focussing then on the instantaneous volatilities, if the simple yet exible functional form discussed proposed in [Reb99] is adopted, i.e. if j (t) is taken to be equal to j (t) = kj (a + b(tj t)) ec(tj t) + d matrix can be calculated in fractions of a second. Given the difculties in estimating reliably and robustly correlation functions (let alone in trying to estimate their possible time dependence), the assumption of time homogeneity for the correlation function is rather appealing. The further assumption that jk = |tj tk | , or as in our particular choice jk = e|tj tk | with = 0.1 , (19) (18)

then the integrals in (17) can be easily carried out analytically (see appendix A), and a whole swaption

is more difcult to defend on purely econometric grounds: it implies, for instance, that the de-correlation between, say, the front and the second forward rate should be the same as the de-correlation between the ninth and the tenth. Luckily, European swaption prices turn out to be relatively insensitive to the details of the correlation function, and this assumption can be shown to produce in most cases prices very little different from those obtained using more complex and realistic correlation functions. 10

5 Empirical Results on European Swaptions

The results and the arguments presented so far indicate that it is indeed plausible that the instantaneous volatility of a swap rate might be evaluated with sufcient precision by calculating the stochastic coefcients {} using the initial yield curve. The ultimate proof of the validity of the procedure, however, is obtained by checking actual European swaption prices. The following test was therefore carried out. rst of all, the instantaneous volatility function described above in equation (18) was used, with parameters chosen as to ensure a realistic and approximately time homogeneous behaviour for the evolution of the term structure of volatilities. This feature was not strictly necessary for the test, but the attempt was made to create as realistic a case study as possible. In particular, the values of the vector k implicitly dened by equation (18) were set to unity, thereby ensuring a time-homogeneous evolution of the term structure of volatilities (see [Reb99] on this point); given this parametrised form for the forward-rate instantaneous volatility, the instantaneous volatility of a given swap was integrated out to the expiry of the chosen Euroepan swaption. The correlation amongst the forward rates was assumed to be given by equation (19). The value of this integral could therefore be evaluated analytically and gave the required approximate implied volatility for the chosen European swaption; with this implied volatility the corresponding approximate Black price was obtained; given the initial yield curve and the chosen instantaneous volatility function for the forward rates, an exact FRA-based BGM/J Monte Carlo evaluation of the chosen European swaption price was carried out. For this evaluation, the same correlation function was used in the estimation of the approximate price, and by retaining as many stochastic driving factors as forward rates in the problem. This meant retaining up to 40 factors for the case studies presented below; the values of the swaps and FRAs that can be obtained as a by-product of the procedure were calculated separately to check against the possible presence of drift biases. The results, (not shown in the tables below) indicated discrepancies from the swap and forward rates always less than a quarter of a basis point with respect to the corresponding reference rate; the price for the European swaption obtained from the simulation, and the corresponding price obtained using the approximate equation (17) was then compared. The results are shown in the tables below. In table 2, we give the discount factors and the resulting prices for at-the-money European swaptions resulting from the different formulas (6) and (15) for a at yield curve at 7% (annually compounded) and a GBP yield curve for August 10th , 2000. The volatilities of the forward rates were modelled according to equation (18) with a = 5%, b = 0.5, c = 1.5, and d = 15%. Correlation was assumed to be as in equation (19) with = 0.1. For each yield curve, all of 40 co-terminal semi-annual at-the-money swaptions with maturity of the nal payment after 20 1/2 years 11

Flat at 7% annually Time [years] 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0 12.5 13.0 13.5 14.0 14.5 15.0 15.5 16.0 16.5 17.0 17.5 18.0 18.5 19.0 19.5 20.0 20.5 discount factor 0.966736 0.934579 0.903492 0.873439 0.844385 0.816298 0.789145 0.762895 0.737519 0.712986 0.689270 0.666342 0.644177 0.622750 0.602035 0.582009 0.562649 0.543934 0.525841 0.508349 0.491440 0.475093 0.459290 0.444012 0.429243 0.414964 0.401161 0.387817 0.374917 0.362446 0.350390 0.338735 0.327467 0.316574 0.306044 0.295864 0.286022 0.276508 0.267311 0.258419 0.249823 Price from eq. (6) 2.572% 3.475% 4.042% 4.427% 4.699% 4.890% 5.022% 5.107% 5.154% 5.169% 5.157% 5.122% 5.065% 4.991% 4.901% 4.796% 4.679% 4.550% 4.411% 4.263% 4.107% 3.943% 3.773% 3.596% 3.414% 3.227% 3.035% 2.840% 2.640% 2.438% 2.232% 2.023% 1.811% 1.597% 1.381% 1.161% 0.938% 0.712% 0.480% 0.242% Price from eq. (15) 2.572% 3.475% 4.042% 4.427% 4.699% 4.890% 5.022% 5.107% 5.154% 5.169% 5.157% 5.122% 5.065% 4.991% 4.901% 4.796% 4.679% 4.550% 4.411% 4.263% 4.107% 3.943% 3.773% 3.596% 3.414% 3.227% 3.035% 2.840% 2.640% 2.438% 2.232% 2.023% 1.811% 1.597% 1.381% 1.161% 0.938% 0.712% 0.480% 0.242% Monte Carlo price 2.569% 3.469% 4.033% 4.416% 4.685% 4.874% 5.004% 5.088% 5.133% 5.147% 5.134% 5.098% 5.042% 4.967% 4.877% 4.773% 4.656% 4.528% 4.390% 4.243% 4.087% 3.925% 3.756% 3.580% 3.400% 3.214% 3.024% 2.830% 2.632% 2.430% 2.226% 2.018% 1.808% 1.594% 1.379% 1.160% 0.938% 0.711% 0.479% 0.242% Vega 0.202% 0.273% 0.318% 0.350% 0.373% 0.389% 0.400% 0.406% 0.409% 0.409% 0.406% 0.402% 0.395% 0.388% 0.378% 0.368% 0.357% 0.345% 0.332% 0.319% 0.305% 0.290% 0.276% 0.261% 0.245% 0.230% 0.214% 0.199% 0.183% 0.167% 0.151% 0.136% 0.120% 0.104% 0.089% 0.074% 0.059% 0.044% 0.029% 0.014% discount factor 0.969514 0.939441 0.909913 0.881024 0.852807 0.825482 0.799100 0.773438 0.749042 0.725408 0.702527 0.680361 0.659402 0.639171 0.619580 0.600668 0.582455 0.564873 0.547888 0.531492 0.515651 0.500360 0.485543 0.471240 0.457861 0.444977 0.432554 0.420575 0.409019 0.397888 0.387341 0.377196 0.367435 0.358056 0.348978 0.340292 0.331614 0.323265 0.315460 0.307945 0.300321

GBP for August 10th , 2000 Price from eq. (6) 2.418% 3.268% 3.798% 4.153% 4.396% 4.561% 4.671% 4.733% 4.765% 4.765% 4.739% 4.689% 4.627% 4.549% 4.454% 4.347% 4.229% 4.101% 3.964% 3.818% 3.665% 3.505% 3.339% 3.167% 2.999% 2.828% 2.653% 2.475% 2.294% 2.112% 1.931% 1.749% 1.565% 1.381% 1.195% 1.008% 0.811% 0.612% 0.415% 0.215% Price from eq. (15) 2.474% 3.344% 3.887% 4.250% 4.497% 4.664% 4.772% 4.832% 4.861% 4.858% 4.827% 4.772% 4.705% 4.622% 4.523% 4.410% 4.287% 4.154% 4.011% 3.860% 3.702% 3.537% 3.365% 3.188% 3.017% 2.842% 2.665% 2.484% 2.301% 2.116% 1.934% 1.751% 1.567% 1.382% 1.195% 1.008% 0.811% 0.611% 0.415% 0.215% Monte Carlo price 2.472% 3.340% 3.880% 4.241% 4.487% 4.653% 4.760% 4.819% 4.848% 4.844% 4.812% 4.757% 4.691% 4.608% 4.509% 4.397% 4.275% 4.142% 4.000% 3.850% 3.692% 3.528% 3.357% 3.181% 3.011% 2.837% 2.660% 2.480% 2.298% 2.113% 1.932% 1.749% 1.565% 1.381% 1.195% 1.008% 0.811% 0.611% 0.415% 0.215% Vega 0.189% 0.254% 0.297% 0.326% 0.347% 0.360% 0.369% 0.374% 0.376% 0.375% 0.372% 0.366% 0.360% 0.352% 0.343% 0.333% 0.322% 0.310% 0.298% 0.285% 0.271% 0.258% 0.244% 0.229% 0.215% 0.201% 0.187% 0.173% 0.159% 0.145% 0.131% 0.117% 0.104% 0.090% 0.077% 0.064% 0.051% 0.038% 0.025% 0.013%

Table 2: The discount factors and results for the two yield curves used in the tests. were calculated. The pricing errors for the different approximations are also shown in gures 5 and 6 for the at and the GBP yield curve, respectively. As expected, the two approximations are identical for a at yield curve. For the GBP yield curve, the approximation (6) diverges from the exact price up to 10 basis points for swaptions of expiry around 3 years. The shape corrected formula (15), however, stays within 1.5 basis points of the correct price which demonstrates how powerful the formula is for the purpose of calibration to European swaptions.

12

Figure 5: The pricing error from the constant-weights approximation given by equation (6) and including the shape correction as in equation (15) for a at yield curve at 7% annual yield.

6 Conclusion
In this article we have derived and analysed a comparatively simple formula for the pricing of European swaptions in the BGM/J framework based on log-normally evolving forward rates. The main result is that a simple summation over weighted FRA/FRA covariances gives a very good approximation for the total equivalent variance incurred by the swap rate. Using the implied volatility equivalent to the total variance in a log-normal option formula, i.e. the Black formula, then sufces to price European swaptions with a remarkable degree of accuracy. We also explained the mechanism responsible for this surprisingly good match between a log-normal pricing formula using an approximate equivalent volatility and a full blown Monte Carlo simulation. The key here was that on average, the weighting coefcients depend only very little on the evolved yield curve for the lower modes of possible deformations, and the higher modes contribute only very little due to their much lower principal components (or eigenvalues) and thus much lower probability of occurring with sufcient amplitude. Finally, we have conducted realistic tests on the validity of the approximation and reported the numerical results in detail.

13

yoo

yoo

n l k i g d omejhfe

oejhe

)yrh@jr

z x w v s q jy@turp

or}{  ~ |

y@

x rw

v iu

is t

q rp

h ig

a ` X @Y W V U S T R Q P H @I G F E C D B A 9 7 @8 6 5 4 2 3 1 0 ( & )' % $ # ! "      

Figure 6: The pricing error from the constant-weights approximation given by equation (6) and including the shape correction as in equation (15) for a GBP yield curve for August 10th , 2000.

The indenite integral of the instantaneous covariance

Given the parametrisation of the instantaneous volatility j (t) of the forward rate fj as in equation (18), and the FRA/FRA correlation (19), the indenite integral of the covariance becomes ij (t)i (t)j (t)dt = e|ti tj | 1 4c3

4ac2 d ec(ttj ) + ec(tti ) + 4c3 d2 t 4bcdec(tti ) c(t ti ) 1 4bcdec(ttj ) c(t tj ) 1 + ec(2tti tj ) 2a2 c2 + 2abc 1 + c(ti + tj 2t) + b2 1 + 2c2 (t ti )(t tj ) + c(ti + tj 2t)

14

)ffj

)ffj

k i h g e fjfd

fjfd

~){nj@fn

fuqnl v t s r p o m

 | z x f}~#{yw

v wu

rs t

rp q

i gh

f ge

x y

c db

V S UT R Q H P@I G F C ED B A 7 9@8 6 5 2 43 1 0 & ()' % $ " #!          

References
[BGM97] A. Brace, D. Gatarek, and M. Musiela. The market model of interest rate dynamics. Mathematical Finance, 7:127155, 1997. [Jam97] F. Jamshidian. Libor and swap market models and measures. Finance and Stochastics, 1:293 330, 1997. [LS95] Ron Levin and Kuljot Singh. What hides behind the smile? Published report, Derivatives Research, J. P. Morgan Securities Inc., New York, Ron Levin: (97-2) 253-6017, Kuljot Singh: (1-212) 648-3104, 1995. [Reb98] [Reb99] Riccardo Rebonato. Interest rate option models. Jon Wiley and Sons, 1998. Riccardo Rebonato. Volatility and Correlation. Jon Wiley and Sons, 1999.

15